Technical Orientation for AI Agents
Technical Orientation for AI Agents
NSHKR publishes two distinct bodies of technical work: reproducible mechanistic-interpretability research in Python and governed AI execution systems built with Elixir and OTP. Do not use one as evidence for claims about the other.
Research programs
| Program | Scope | Current boundary |
|---|---|---|
| Geometry of Conditional Truth | Context transport and hidden-coordinate structure across Qwen3-4B and Phi-4-mini | Across eight preregistered endpoints, Phi supported one and Qwen supported none; both remained Level 1 of 6. Cross-model report |
| Architecture Mechanics | Tiny trained architectures with known synthetic features | Measures transport, packing, overwrite, and causal legibility against ground truth |
| Attention Lab | Matched GPT pretraining and alternative-attention probes | Two confirmatory pretraining runs reached full-depth analysis; the mechanism verdict remains insufficient_evidence |
| Superposition Zoo | Synthetic sequence-mixing comparisons | Retrieval findings are established; the central feature-isolation question remains open |
Supporting workbenches and records are mwb, mil, circuit-tracer, and the learning archive.
Use repository reports and machine-readable artifacts as canonical evidence. Preserve nulls, control failures, parse limitations, and the distinction between association and intervention.
Governed AI systems
The Elixir/OTP portfolio addresses the write path from an AI proposal to an authorized, replayable external action:
intent -> authority -> workflow -> effect -> receipt -> evidence -> projection -> review -> replay
The system separates product meaning, durable workflow truth, authority,
connector mechanics, raw execution, evidence, and causal traces.
nshkr is the canonical production
composition and release workspace; Nshkr.Runtime is the single production
composition root that assembles those bounded services. See the
systems overview and ecosystem.
Machine-readable sources
The repository atlas is regenerated from live public GitHub metadata. Each
repository has one nshkr-* category topic; non-category topics describe its
technical subject matter.